search

dyabel / detpro

Apache License 2.0
174 stars 28 forks source link

Training RPN based on the provided code. Rather low performance. #38

Open ustcjinggg opened 1 year ago

ustcjinggg commented 1 year ago

Hi, Thanks for sharing the code and model. I've tried the code of RPN you provided and work on building a more generalized rpn. However, I have the following issues and would appreciate your help : I test the epoch20.pth you provide on lvis_1 dataset. I get AR@100=36.5, far from the AR given by ViLD paper, Table 1: AR@100=39.3. Why does this happen? I understand that ViLD tests on the lvis novel data and you test the whole data. Shouldn't the epoch_20.pth perform much better than the ViLD rpn? I don't get the same performance when training RPN using your code. I use the config rpn_r101_fpn_1x_lvis.py (training 12 epoch, no 20epoch config is provided) and only get AR@100=9.5. I follow the mmdet instructions and simply use bash tools/dist_tran.sh configs/rpn/rpn_r101_fpn_1x_lvis.py 2 to train the RPN, why do I get such low performance? I can only test the RPN model using proposal_fast mode. When I use bash tools/dist_test.sh configs/rpn/rpn_r101_fpn_1x_lvis.py models/epoch_20.pth 1 --out work_dirs/result.pkl --eval proposal , I get all zero results. image

ustcjinggg commented 1 year ago

image

Kegard commented 1 year ago

hello, I want to know how do you Train RPN on based? I cannot found the readme about this, just some code.